Keep work moving.
Maintain state across long sessions. Coordinate workers, pause for input, and recover from recorded evidence.
Sessions & recoveryOpen source · Python · Apache 2.0
Build domain-specific
long-horizon agents.
Cayu is an open-source Python framework for building and running domain-specific long-horizon agents. Compose a harness around your tools, knowledge, and business rules—with a durable runtime built in.
pip install cayuHow Cayu works
Build the agent.
Shape how it works.
Cayu brings the building blocks, harness components, and runtime together in one Python framework. You supply the domain knowledge and define how the agent does its job.
| Concept | Its role in Cayu |
|---|---|
| Framework | The Python structure, APIs, defaults, and components you build with. |
| Harness | The system you assemble with Cayu: context, tools, memory, policies, domain logic, and verification. |
| Runtime | The integrated execution machinery: sessions, state, workers, recovery, approvals, and budgets. |
| Agent | The resulting product that performs work in your domain. |
An agent built with CayuA product that performs work in your domain
Cayu frameworkPython APIs, defaults, and components to build with
Your domain-specific harnessAssembled and customized with Cayu
- Context
- Tools
- Policies
- Domain logic
- Memory
- Verification & evals
- 01
Model
Propose the next step
Your choice of provider - 02
Tools
Check permission, then act
Your systems and functions - 03
Result
Collect and verify evidence
Context for what comes next
Continue toward the goal Wait for input Finish
Cayu runtimeIntegrated execution, built in
Durable sessions & stateWorkers & recoveryApprovals & budgetsReplay & observability
Long-horizon work
A clear goal.
Hours of work to get there.
“Process these 5,000 documents” has a finish line. Getting there may involve thousands of calls, new evidence, a human decision, and a restart.
Document review
Hour 3
One goal. Thousands of decisions.
Review 5,000 documents, extract the required fields, and flag exceptions. Your harness defines the checks. Cayu records the execution.
The goal stays bounded as the execution grows.
Documents reviewed1,240 / 5,000
Estimated model cost$18.60 / $100 budget
Document batch completed
Results and session progress recorded
Next
Continue reviewing the remaining documents.
Hour 17
Ask a person. Keep the context.
Sixteen documents need a reviewer’s decision. The session waits with its evidence and progress intact, ready for the next input.
Waiting for a person is part of the workflow.
Documents reviewed3,240 / 5,000
Estimated model cost$42.80 / $100 budget
Reviewer input requested
16 exceptions attached to the session
Next
Waiting for the reviewer to confirm classifications.
Hour 20
The worker stops. The work is recorded.
After the reviewer responds, processing continues. Then the worker exits. Durable records preserve what is known; uncertain actions remain explicit.
A stopped process should not erase the history.
Documents reviewed3,920 / 5,000
Estimated model cost$51.40 / $100 budget
Worker connection lost
Last confirmed batch preserved in durable storage
Next
Inspect the recorded state before continuing.
Hour 20 · resumed
Continue from the recorded state.
A replacement worker loads the session. In this example, the completed batches are confirmed and the remaining documents can be processed.
Ambiguous external effects require reconciliation before retrying.
Documents reviewed3,920 / 5,000
Estimated model cost$51.40 / $100 budget
Session recovered
Confirmed results restored · 1,080 documents remaining
Next
Resume from the last confirmed batch.
Hour 26
Finish against your definition of done.
All documents are accounted for, the reviewer’s decisions are applied, and your harness’s completion checks pass. The result is ready with its supporting evidence.
Completion is checked against criteria you define.
Documents reviewed5,000 / 5,000
Estimated model cost$64.10 / $100 budget
Completion checks passed
5,000 documents accounted for · 16 exceptions resolved
Next
Results, decisions, and execution history are available.
This example shows the lifecycle you can build with Cayu. Your harness defines the business rules and completion checks.
Start building
Python you can
make your own.
Start with a project that has a place for your prompts, tools, policies, memory, and evals. Bring the libraries and domain code your team already uses.
Read the quickstartPython 3.11+ · Run locally or on your infrastructure
pip install cayu
cayu new support-agent
cd support-agentcayu check --fail-on warning --json
pytest
cayu eval runConfigure your model provider in configuration/settings.py or CAYU_PROVIDER first.
python run.py --message "Review this change."Build with Codex, Claude Code, or Cursor
Open the generated project in your development tool, or start with this prompt and build and refine your agent together.
Install cayu and create a document review agent. Use cayu new to scaffold the project. Define the required fields, flag exceptions for human review, and add completion checks and evals. Ask me about the document format and model provider before configuring them.
What Cayu handles
The foundation behind the work.
Keep actions under control.
Define which tools an agent may use, when approval is required, and how much a run may spend.
Permissions & budgetsUnderstand and improve.
Inspect the work, trace model costs, and evaluate behavior using the path the agent took and the result it produced.
Observability & evalsBuild, then operate
Your agent.
Your way to run it.
Build your agent and its domain-specific harness with Cayu. Run it on your own infrastructure, or use Cloud as your team’s shared workspace.
Why we built it
Built from the work.
We kept rebuilding the same infrastructure around agents for finance, accounting, and enterprise operations. The complexity grew when the work had to continue across days, decisions, and failures.
We brought those building blocks and the runtime together in Cayu. Our background in ML and industrial systems shapes the approach: domain knowledge, verification, and the system around the model matter.
Meet CayuA few useful answers
What is Cayu?
Cayu is an open-source Python framework for building and running domain-specific long-horizon agents. It provides components and defaults to assemble your agent’s harness, with an integrated runtime for durable, controlled execution.
What is an agent harness?
A harness is the system around the model: context, tools, policies, domain logic, memory, and verification. You assemble and customize it with Cayu to define how your agent does its job.
When does Cayu become useful?
When work spans many decisions, interacts with real systems, waits for people, needs to survive interruptions, or operates under cost and permission constraints. The goal can be bounded even when execution takes hours or days.
Can I choose the model and control the loop?
Yes. Cayu provides defaults and replaceable Python contracts for models, context, tools, execution, and storage. Supported provider integrations include OpenAI, Anthropic, Amazon Bedrock, Google Vertex AI, and OpenAI-compatible APIs. You can implement your own provider contract.
Do I need Cayu Cloud?
No. Agents built with Cayu run on infrastructure you choose, with a packaged operator dashboard included. Cayu Cloud is an optional shared workspace for building, deploying, operating, and improving your agents.
Is Cayu open source?
Yes. Cayu is licensed under Apache 2.0. You can inspect, modify, and self-host it. We use Cayu in enterprise deployments; the open-source project is early, and we welcome feedback from teams building domain-specific long-horizon agents.
Build for the long run
Bring your domain.
Build your agent.
What does your agent need to do?
Join the conversation ↗